Breakthrough in passive non-line-of-sight imaging: UMA model enables high-quality multi-object reconstruction

The team proposed a new passive NLOS imaging model UNet MultiAttention (UMA). The UMA model utilizes nested UNet architecture and joint attention mechanisms (channel, spatial, and self attention) for feature extraction and fusion. UMA dynamically prioritizes meaningful spatial and channel features to accurately reconstruct complex environments. In addition, a joint loss function combining BCE and MSE is used to supervise multiple reconstruction branches at different semantic levels. Conduct experiments on an infrared multi-target NLOS dataset (scene ranges from 1 to 3 hidden objects). The results showed an average improvement of 4.8% on PSNR and 2.5%on SSIM compared to the state-of-the-art NLOS passive method. Our future work will focus on exploring how to use reconstruction results for NLOS target recognition.